What it is. Grown organically from three capabilities — caching, column selection (pick the column family at its best input anchor), and usage-driven dynamic optimal caching — the Cache(r) is the sound version of the role OLAP called the aggregate navigator: answer a query from an already-materialized column whenever a lawful path exists. The difference from thirty years of prior art is the proof obligation: a cached column at anchor A serves a request at anchor A′ iff the algebra certifies the reduction A→A′ — same criterion that gates a fresh query. A cache hit is a theorem application. The same anchors that block illegal queries certify legal shortcuts: the governance metadata pays for itself in round-trips saved.
Rulings (2026-07-12)
- Name: Cache(r) — sibling to Combine(r); role-named per the labeling discipline.
- Substrate: Polars — same dual-role rule as the Combiner: the library is incidental; the role is what is fixed.
- Freshness & hot/cold consistency: general caching obligations, nothing Columna-specific — discharged through the Manifold's existing contract component (data attestation, universe definitions). The cache inherits its obligations from the contract; it does not invent doctrine. Natural test expression: a hot/cold parity suite — cache-hit and cold computation produce identical wire responses modulo freshness metadata.
- Status: design-stage, to be purpose-built — including the telemetry posture (usage-driven optimization consumes query logs; scope, retention, and locality of usage stats are part of the deliberate build-out, stated before the role is marketed).
Optimization space (recorded)
The candidate set is structured by the Manifold — (family, anchor) pairs — and pruned by legality: path-independent reductions at coarse anchors have rich legal descendant cones (load-bearing cache entries); a non-commuting reduction's result can serve only itself (a leaf). Cache at the anchors with the richest lawful descendant cones, weighted by observed usage.
The layered Cache(r) (large installations serving many Manifolds)
Per-Manifold engines remain the rule; atop them, shared cache layers — enterprise, department, project — are possible because of the grammar: a cached column's identity is its semantic identity (family, input anchor, universe, attestation lineage), checkable by the algebra, so sharing is a theorem about column identity rather than a trust arrangement between teams. Names cannot fragment the cache — a metadata-independence dividend: communicative metadata varies freely while cache keys ride only the load-bearing kind. Promotion into a shared layer is gated by the same vetting/certification machinery that gates everything else.
Governance consequences (recorded, to shape the build-out):
- The shareable cache is exactly as large as the shared grammar — sharing presupposes common (or explicitly mapped) families and universes at the sharing layer; the enterprise cache's performance payoff becomes the standing economic incentive for semantic convergence across departments.
- Layer visibility follows data scope; invalidation flows down. A Manifold reads a shared-layer column only where its scope and universe legitimately contain it; the cache hierarchy respects the scope lattice as well as the reduction lattice. Attestation changes at a shared layer propagate invalidation to every inheriting layer — riding the contract machinery, with the propagation paths made explicit.
Origin note. The role was not designed from the OLAP angle; it precipitated out of the algebra — sound by construction — because the grammar was already carrying exactly the information navigation needs. The container world bolted navigation on and prayed; in the substance world it condenses for free.